def _Build_GenerativeNetwork(self, g_input):
generate_weights = _initialize_generate_variables()
layer0 = tf.nn.relu(tf.matmul(g_input, generate_weights['w0']) + generate_weights['b0'])
drop_layer0 = tf.nn.dropout(layer0, self.keep)
layer1 = tf.nn.relu(tf.matmul(layer0, generate_weights['w1']) + generate_weights['b1'])
drop_layer1 = tf.dropout(layer1, self.keep)
output_layer = tf.nn.relu(tf.matmul(layer1, generate_weights['w2']) + generate_weights['b2'])
drop_output_layer = tf.dropout(output_layer, self.keep)
return drop_output_layer
def _Build_AdversarialNetwork(self, a_input):
adversary_weights = _initialize_adversary_variables()
layer0 = tf.nn.relu(tf.matmul(a_input, adversary_weights['w0']) + adversary_weights['b0'])
drop_layer0 = tf.dropout(layer0, self.keep)
layer1 = tf.nn.relu(tf.matmul(layer0, adversary_weights['w1']) + adversary_weights['b1'])
prob = tf.nn.softmax(tf.matmul(layer1, adversary_weights['w2']) + adversary_weights['b2'])
return prob
def fm(self, x):
with tf.variable_scope("linear_layer"):
w = tf.get_variable("w",
shape=[self.feature_size, 1],
initializer=tf.truncated_normal_initializer(
mean=0, stddev=1e-2))
b = tf.get_variable("b",
shape=[1],
initializer=tf.zeros_initializer())
self.linear_output = tf.matmul(x, w) + b #(batch, field_size, 1)
self.linear_output = tf.dropout(self.linear_output,
self.fm_keep_prob)
with tf.variable_scope("interaction_layer"):
# sum square part
sum_sq_part = tf.square(tf.reduce_sum(self.embeddings,
1)) #(batch, embedding_size)
# square sum part
sq_sum_part = tf.reduce_sum(tf.square(self.embeddings),
1) #(batch, embedding_size)
self.interaction_output = 0.5 * tf.subtract(
sum_sq_part, sq_sum_part)
self.interaction_output = tf.nn.dropout(self.interaction_output,
self.fm_keep_prob)
return (self.linear_output, self.interaction_output)
def call(self, inputs, **kwargs):
main_input, embedding_matrix = inputs
input_shape_tensor = tf.shape(main_input)
last_input_dim = tf.int_shape(main_input)[-1]
emb_input_dim, emb_output_dim = tf.int_shape(embedding_matrix)
projected = tf.dot(tf.reshape(main_input, (-1, last_input_dim)),
self.projection)
if self.add_biases:
projected = tf.bias_add(projected,
self.biases,
data_format='channels_last')
if 0 < self.projection_dropout < 1:
projected = tf.in_train_phase(
lambda: tf.dropout(projected, self.projection_dropout),
projected,
training=kwargs.get('training'))
attention = tf.dot(projected, tf.transpose(embedding_matrix))
if self.scaled_attention:
# scaled dot-product attention, described in
# "Attention is all you need" (https://arxiv.org/abs/1706.03762)
sqrt_d = tf.constant(math.sqrt(emb_output_dim), dtype=tf.floatx())
attention = attention / sqrt_d
result = tf.reshape(
self.activation(attention),
(input_shape_tensor[0], input_shape_tensor[1], emb_input_dim))
return result
def neural_net_model(input_size):
# Input Layer
net = tf.input_layer(shape=[None, input_size], name='input')
# Hidden Layers
net = tf.fully_connected(net, 128, activation='relu', name="hlayer1")
net = tf.dropout(net, 0.8)
net = tf.fully_connected(net, 256, activation='relu', name="hlayer2")
net = tf.dropout(net, 0.8)
net = tf.fully_connected(net, 512, activation='relu', name="hlayer3")
net = tf.dropout(net, 0.8)
net = tf.fully_connected(net, 256, activation='relu', name="hlayer4")
net = tf.dropout(net, 0.8)
net = tf.fully_connected(net, 128, activation='relu', name="hlayer5")
net = tf.dropout(net, 0.8)
# Output layer
net = tf.fully_connected(net, 2, activation='softmax', name="out")
net = tf.regression(net, learning_rate = LR)
model = tf.DNN(net, tensorboard_dir='log')
return model
def call(self, x, adj, training=True):
if self.use_bn:
x = self.bn1(x, training=training)
if training:
x = tf.dropout(x, self.input_droprate)
x = tf.nn.relu(self.gc1(x, adj))
if self.use_bn:
x = self.bn2(x, training=training)
if training:
x = tf.nn.dropout(x, self.hidden_droprate)
x = self.gc2(x, adj)
return x
def build_model(self):
with tf.variable_scope("SeqClassificationRNN"):
# 训练数据
self.inputs = tf.placeholder(tf.int32,
shape=[None, self.max_length])
self.seqlen = tf.placeholder(tf.int32, shape=[None])
# 训练标签数据
self.labels = tf.placeholder(tf.int32, shape=[None])
onehot = tf.one_hot(self.labels, self.class_num)
# dropout
self.keep_prob = tf.placeholder(tf.float32)
embedding_matrix = tf.Variable(
tf.truncated_normal((self.vocab_size, self.embed_dim),
stddev=0.01))
embedding = tf.nn.embedding_lookup(embedding_matrix, self.inputs)
cell = tf.nn.rnn_cell.MultiRNNCell([
self.cell(self.hidden_dim, state_is_tuple=True)
for _ in range(self.layer_num)
])
outputs, state = tf.nn.dynamic_rnn(cell,
embedding,
self.seqlen,
dtype=tf.float32,
swap_memory=True)
pooling = tf.reduce_sum(outputs, 1) / self.seqlen
dropout = tf.dropout(pooling, keep_prob=self.keep_prob)
lr_W = tf.Variable(
tf.truncated_normal((self.hidden_dim, self.class_num),
stddev=0.1))
lr_b = tf.Variable(tf.zeros((1, self.class_num)))
pred = tf.matmul(dropout, lr_W) + lr_b
# 定义交叉熵损失函数
self.loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(pred, onehot))
def _build_model(self):
"""
It does not want arguments because it is going to pick whatever it needs.
"""
start_trainable_variables = tf.trainable_variables()
net = self.inputs
for layer_dim in layer_dims:
net = tf.layers.dense(inputs=net, units=layer_dim, activation=tf.nn.leaky_relu)
# Never dropout for first or last layer.
if not self.test_phase:
net = tf.dropout(net, 0.5)
# last layer back to the original input dimension
self.outputs = tf.layers.dense(inputs=net, units=self.output_dim, activation=tf.identity)
end_trainable_variables = tf.trainable_variables()
self.L_params = [ param for param in end_trainable_variables if param not in start_trainable_variables ]
def build_model(self, data_sources: Dict[str, BaseDataSource], mode: str):
"""Build model."""
data_source = next(iter(data_sources.values()))
input_tensors = data_source.output_tensors
x = input_tensors['left-eye']
batch_size = 32
# Trainable parameters should be specified within a known `tf.variable_scope`.
# This tag is later used to specify the `learning_schedule` which describes when to train
# which part of the network and with which learning rate.
#
# This network has two scopes, 'conv' and 'fc'. Though in practise it makes little sense to
# train the two parts separately, this is possible.
with tf.variable_scope('conv'):
with tf.variable_scope('conv1'):
x = tf.pad(x, [[0, 0], [0, 0], [0, 1], [0, 1]], "constant")
x = tf.layers.conv2d(x, filters=64, kernel_size=3, strides=2,
padding='valid', data_format='channels_first')
# self.summary.filters('filters', x)
x = tf.nn.relu(x)
# self.summary.feature_maps('features', x, data_format='channels_first')
x = tf.pad(x, [[0, 0], [0, 0], [0, 1], [0, 1]], "constant")
x = tf.layers.dropout(x, rate = 0.1, noise_shape = (batch_size, 128, 1, 1), training=False)
x = tf.layers.conv2d(x, filters=128, kernel_size=3, strides=1,
padding='valid', data_format='channels_first')
x = tf.nn.relu(x)
x = tf.layers.max_pooling2d(x, pool_size=3, strides=2, padding='valid', data_format='channels_first')
# self.summary.feature_maps('features', x, data_format='channels_first')
with tf.variable_scope('conv2'):
x = tf.pad(x, [[0, 0], [0, 0], [1, 1], [1, 1]], "constant")
x = tf.layers.conv2d(x, filters=256, kernel_size=3, strides=1,
padding='valid', data_format='channels_first')
# self.summary.feature_maps('features', x, data_format='channels_first')
x = tf.nn.relu(x)
#x = tf.pad(x, [[0, 0], [0, 0], [1, 1], [1, 1]], "constant")
# x = tf.layers.dropout (x, rate=0.1, noise_shape=(batch_size, 512,1,1), training=False)
# x = tf.layers.conv2d(x, filters=256, kernel_size=5, strides=2,
# padding='same', data_format='channels_first')
# self.summary.feature_maps('features', x, data_format='channels_first')
# x = tf.nn.relu(x)
x = tf.layers.max_pooling2d(x, pool_size=3, strides=2, padding='valid', data_format='channels_first')
with tf.variable_scope('conv3'):
x = tf.pad(x, [[0, 0], [0, 0], [1, 1], [1, 1]], "constant")
x = tf.layers.conv2d(x, filters=512, kernel_size=3, strides=1,
padding='valid', data_format='channels_first')
# self.summary.feature_maps('features', x, data_format='channels_first')
x = tf.pad(x, [[0, 0], [0, 0], [1, 1], [1, 1]], "constant")
x = tf.layers.dropout (x, rate=0.1, noise_shape=(batch_size, 512, 1, 1), training=False)
x = tf.layers.conv2d(x, filters=512, kernel_size=3, strides=1,
padding='valid', data_format='channels_first')
# self.summary.feature_maps('features', x, data_format='channels_first')
x = tf.pad(x, [[0, 0], [0, 0], [1, 1], [1, 1]], "constant")
x = tf.layers.conv2d(x, filters=512, kernel_size=3, strides=1,
padding='valid', data_format='channels_first')
# self.summary.feature_maps('features', x, data_format='channels_first')
x = tf.pad(x, [[0, 0], [0, 0], [1, 1], [1, 1]], "constant")
x = tf.layers.dropout(x, rate=0.1, noise_shape=(batch_size, 512, 1, 1), training=False)
x = tf.layers.conv2d(x, filters=512, kernel_size=3, strides=1, padding='valid', data_format='channels_first')
x = tf.layers.max_pooling2d(x, pool_size=3, strides=2,
padding='same', data_format='channels_first')
with tf.variable_scope('conv4'):
x = tf.pad(x, [[0, 0], [0, 0], [1, 1], [1, 1]], "constant")
x = tf.dropout(x, rate=0.1, noise_shape=(batch_size, 512, 1, 1), training=False)
x = tf.layers.conv2d(x, filters=512, kernel_size=3, strides=1,
padding='valid', data_format='channels_first')
x = tf.pad(x, [[0, 0], [0, 0], [1, 1], [1, 1]], "constant")
# x = tf.layers.dropout(x, rate=0.1, noise_shape=(batch_size, 1024, 1, 1), training=False)
x = tf.layers.conv2d(x, filters=256, kernel_size=3, strides=2,
padding='same', data_format='channels_first')
#x = tf.pad(x, [[0, 0], [0, 0], [1, 1], [1, 1]], "constant")
# x = tf.layers.dropout(x, rate=0.1, noise_shape=(batch_size, 1024, 1, 1), training=False)
x = tf.layers.conv2d(x, filters=256, kernel_size=3, strides=2,
padding='same', data_format='channels_first')
x = tf.layers.max_pooling2d(x, pool_size=3, strides=2, padding='same', data_format='channels_first')
with tf.variable_scope('conv5'):
#x = tf.pad(x, [[0, 0], [0, 0], [1, 1], [1, 1]], "constant")
x = tf.layers.conv2d(x, filters=512, kernel_size=3, strides=2,
padding='same', data_format='channels_first')
x = tf.pad(x, [[0, 0], [0, 0], [1, 1], [1, 1]], "constant")
# x = tf.layers.dropout(x, rate=0.1, noise_shape=(batch_size, 1024, 1, 1), training=False)
x = tf.layers.conv2d(x, filters=512, kernel_size=3, strides=2,
padding='valid', data_format='channels_first')
x = tf.pad(x, [[0, 0], [0, 0], [1, 1], [1, 1]], "constant")
x = tf.layers.dropout(x, rate=0.1, noise_shape=(batch_size, 512, 1, 1), training=False)
x = tf.layers.conv2d(x, filters=512, kernel_size=3, strides=1,
padding='valid', data_format='channels_first')
x = tf.layers.max_pooling2d(x, pool_size=3, strides=2, padding='same', data_format='channels_first')
with tf.variable_scope('fc'):
# Flatten the 50 feature maps down to one vector
x = tf.contrib.layers.flatten(x)
# FC layer
x = tf.layers.dense(x, units=4096, activation='relu', name='fc5')
x = tf.layers.dense(x, units=4096, activation='relu', name='fc6')
x = tf.layers.dense(x, units=1024, activation='softmax', name='fc7')
self.summary.histogram('fc7/activations', x)
# Directly regress two polar angles for gaze direction
x = tf.layers.dense(x, units=2, name='fc8')
self.summary.histogram('fc8/activations', x)
# Define outputs
loss_terms = {}
metrics = {}
if 'gaze' in input_tensors:
y = input_tensors['gaze']
with tf.variable_scope('mse'): # To optimize
loss_terms['gaze_mse'] = tf.reduce_mean(tf.squared_difference(x, y))
with tf.variable_scope('ang'): # To evaluate in addition to loss terms
metrics['gaze_angular'] = util.gaze.tensorflow_angular_error_from_pitchyaw(x, y)
return {'gaze': x}, loss_terms, metrics
def dropped_softmax():
return tf.dropout(attention_softmax, self.dropout)
# Load the weights and bias #Since tf.train.Saver.restore() sets all the TensorFlow Variables, you don't need to call tf.global_variables_initializer(). #it will restore all the previously saved variables saver.restore(sess, save_file) #when tf try to load the variables it will load variable with the matching name , if #no name was assigned to the variable when assigned ,it will assign the nanme type_number and would #potentially confused when load back the stored variable ,due to mis match of the name weights = tf.Variable(tf.truncated_normal([2, 3]), name='weights_0') #return the index of the max value of that axies, axis = 0 compare rows, axis = 1 compare columns tf.argmax(input, axis=None, name=None, dimension=None) #drop out function #n order to compensate for dropped units, tf.nn.dropout() multiplies all units that are kept (i.e. not dropped) by 1/keep_prob tf.dropout(hidden_layer,keep_prob) #np.pad a = [1, 2, 3, 4, 5] np.pad(a, (2,3), 'constant', constant_values=(4, 6)) array([4, 4, 1, 2, 3, 4, 5, 6, 6, 6]) # ---------------------------------------------------------------------------------------- #convolution network #stride is the stride for each dimention of the input tf.nn.conv2d(X,W1, strides = [1,s,s,1], padding = 'SAME') #tf.nn.max_pool(A, ksize = [1,f,f,1], strides = [1,s,s,1], padding = 'SAME'): given an input A, #this function uses a window of size (f, f) and strides of size (s, s) #to carry out max pooling over each window. tf.nn.max_pool(A, ksize = [1,f,f,1], strides = [1,s,s,1], padding = 'SAME'):
def SublayerConnection(self, x, keep_rate):
return batch_norm(
(x + tf.dropout(self.model(x, x, x), keep_rate=keep_rate)))
def MultiRNN(x,
BATCH_SIZE,
seq,
NUM_CLASSES,
NUM_LSTM,
NUM_HIDDEN,
OUTPUT_KEEP_PROB,
NUM_MLP,
NUM_NEURON,
training=True):
"""model a LDNN Network,
Args:
x: feature, shape = [batch_size, time_length,160]
Returns:
original_out: prediction
update_op: resume state from previous state
reset_op: not use in train, only for validation to reset zero
"""
with tf.variable_scope('lstm', initializer=tf.orthogonal_initializer()):
"""Runs the forward step for the RNN model.
Args:
inputs: `3-D` tensor with shape `[time_len, batch_size, input_size]`.
initial_state: a tuple of tensor(s) of shape
`[num_layers * num_dirs, batch_size, num_units]`. If not provided, use
zero initial states. The tuple size is 2 for LSTM and 1 for other RNNs.
training: whether this operation will be used in training or inference.
Returns:
output: a tensor of shape `[time_len, batch_size, num_dirs * num_units]`.
It is a `concat([fwd_output, bak_output], axis=2)`.
output_states: a tuple of tensor(s) of the same shape and structure as
`initial_state`.
"""
mlstm_cell = tf.contrib.cudnn_rnn.CudnnLSTM(NUM_LSTM, NUM_HIDDEN)
# dropout connect
for w in mlstm_cell.variables:
tf.assign(w, tf.dropout(w, OUTPUT_KEEP_PROB))
states = get_state_variables(NUM_LSTM, BATCH_SIZE, NUM_HIDDEN)
# get shape, and add inputs input_shape attributes
# batch_x_shape = tf.shape(x)
x = tf.reshape(x, [BATCH_SIZE, -1, 160])
# batch_major -> time_length_major
inputs = tf.transpose(x, [1, 0, 2])
outputs, new_states = mlstm_cell(inputs, states, training=training)
# time_length_major -> batch_major
outputs = tf.transpose(outputs, [1, 0, 2])
update_op = get_state_update_op(states, new_states, NUM_LSTM)
# TODO: reset the state to zero or the final state og training??? Now is zero.
reset_op = get_state_reset_op(states, BATCH_SIZE, NUM_LSTM, NUM_HIDDEN)
outputs = tf.reshape(outputs, [-1, NUM_HIDDEN])
with tf.variable_scope('mlp'):
weights = {
'out':
tf.get_variable(
'out',
shape=[NUM_HIDDEN, NUM_CLASSES],
initializer=tf.contrib.layers.xavier_initializer()),
'h1':
tf.get_variable(
'h1',
shape=[NUM_HIDDEN, NUM_NEURON],
initializer=tf.contrib.layers.xavier_initializer()),
'h2':
tf.get_variable(
'h2',
shape=[NUM_NEURON, NUM_NEURON],
initializer=tf.contrib.layers.xavier_initializer()),
'h3':
tf.get_variable(
'h3',
shape=[NUM_NEURON, NUM_NEURON],
initializer=tf.contrib.layers.xavier_initializer()),
'mlpout':
tf.get_variable('mlpout',
shape=[NUM_NEURON, NUM_CLASSES],
initializer=tf.contrib.layers.xavier_initializer())
}
if NUM_MLP == 0:
top = tf.matmul(outputs, weights['out'])
original_out = tf.reshape(top, [BATCH_SIZE, -1, NUM_CLASSES])
return original_out, update_op, reset_op
elif NUM_MLP == 1:
l1 = tf.nn.dropout(tf.matmul(outputs, weights['h1']),
keep_prob=OUTPUT_KEEP_PROB)
l1 = tf.nn.relu(l1)
top = tf.matmul(l1, weights['mlpout'])
original_out = tf.reshape(top, [BATCH_SIZE, -1, NUM_CLASSES])
return original_out, update_op, reset_op
elif NUM_MLP == 2:
l1 = tf.nn.dropout(tf.matmul(outputs, weights['h1']),
keep_prob=OUTPUT_KEEP_PROB)
l1 = tf.nn.relu(l1)
l2 = tf.nn.dropout(tf.matmul(l1, weights['h2']),
keep_prob=OUTPUT_KEEP_PROB)
l2 = tf.nn.relu(l2)
top = tf.matmul(l2, weights['mlpout'])
original_out = tf.reshape(top, [BATCH_SIZE, -1, NUM_CLASSES])
return original_out, update_op, reset_op
elif NUM_MLP == 3:
l1 = tf.nn.dropout(tf.matmul(outputs, weights['h1']),
keep_prob=OUTPUT_KEEP_PROB)
l1 = tf.nn.relu(l1)
l2 = tf.nn.dropout(tf.matmul(l1, weights['h2']),
keep_prob=OUTPUT_KEEP_PROB)
l2 = tf.nn.relu(l2)
l3 = tf.nn.dropout(tf.matmul(l2, weights['h3']),
keep_prob=OUTPUT_KEEP_PROB)
l3 = tf.nn.relu(l3)
top = tf.matmul(l3, weights['mlpout'])
original_out = tf.reshape(top, [BATCH_SIZE, -1, NUM_CLASSES])
return original_out, update_op, reset_op
def __init__(self,
sequence_length,
num_classes,
vocab_size,
embedding_size,
filter_sizes,
num_filters,
l2_reg_lambda=0.0):
"""
:param sequence_length: 句子的长度，固定同意长度
:param num_classes: 输出层的神经元个数
:param vocab_size: 词典的总数用于输入，[vocabulary大小, embedding大小]
:param embedding_size: 嵌入层的维度
:param filter_sizes: 卷积层需要覆盖的单次数[2,3,4]
:param num_filters: 每个尺寸的滤波器的个数， 例如100个
:param l2_reg_lambda: L2正则化
"""
# 输入输出和dropout的占位符
self.input_x = tf.placeholder(tf.int32,
shape=[None, sqeuence_length],
name='input_x')
self.input_y = tf.placeholder(tf.float32,
shape=[None, num_classes],
name='input_y')
self.droupout_keep_prob = tf.placeholder(tf.float32,
name='dropout_keep_prob')
# L2正则化
l2_loss = tf.constant(0.0)
# 嵌入层
with tf.device('/cpu:0'), tf.name_scope('embedding'):
self.W = tf.Variable(tf.random_uniform(
[vocab_size, embedding_size], -1.0, 1.0),
name='W')
# 输入[输入张量， 张量对应的索引]
# 输出[None, sequence_length，embedding_size]
self.embedded_chars = tf.nn.embedding_lookup(self.W, self.input_x)
# 卷积需要四维tensor
# expand_dim和reshape都可以改变维度，但是在构建具体的图时，如果没有具体的值，使用reshape则会报错
self.embedded_chars_expanded = tf.expand_dims(
self.embedded_chars, -1)
# 为每一个filter size构建卷积层和最大池化层
pooled_outputs = []
for i, filter_size in enumerate(filter_size):
with tf.name_scope('conv-maxpool-%s' % filter_size):
# 卷积层
# 卷积核，[卷积核的高度和宽度，通道个数，卷积核个数]
filter_shape = [filter_size, embedding_size, 1, num_filters]
W = tf.Variable(tf.truncated_normal(filter_shape, stddev=0.1),
name='W')
b = tf.Variable(tf.constant(0.1, shape=[num_filter]), name='b')
# padding: SAME 表示用0来填充，VALID用来表示不填充
# strides: [batch, height,width,channels]，batch和channels为1
conv = tf.nn.conv2d(self.embedded_chars_expanded,
W,
strides=[1, 1, 1, 1],
padding='VALID',
name='relu')
# relu
h = tf.nn.relu(
tf.nn.max_pool(
h,
ksize=[1, sequence_length - filter_size + 1, 1, 1],
strides=[1, 1, 1, 1],
padding='VALID',
name='pool'))
pooled_outputs.append(pooled)
# 组合所有的池化层特征
num_filter_total = num_filters * len(filter_sizes)
# 在pooled_outputs 的最后一个维度上连接
self.h_pool = tf.concat(pooled_outputs, 3)
self.h_pool_flat = tf.reshape(self.h_pool, [-1, num_filter_total])
# 增加dropout
with tf.name_scope('dropout'):
self.h_drop = tf.dropout(self.h_pool_flat, self.dropout_keep_prob)
# 最后unnormalized scores and predictions
with tf.name_scope('output'):
W = tf.get_variable(
'W',
shape=[num_filter_total, num_classes],
initializer=tf.contrib.layers.xavier_initializer())
b = tf.Variable(tf.constant(0.1, shape=[num_classes]), name='b')
l2_loss += tf.nn.l2_loss(W)
l2_loss += tf.nn.l2_loss(b)
self.scores = tf.nn.xw_plus_b(self.h_drop, W, b, name='scores')
self.prediction = tf.argmax(self.scores, 1, name='predictions')
# 计算 平均 cross-entropy loss
with tf.name_scope('loss'):
losses = tf.nn.softmax_cross_entopy_with_logit_v2(
logits=self.scores, labels=self.input_y)
#l2 正则化
self.loss = tf.reduce_mean(losses) + l2_reg_lambda * l2_loss
# accuracy
with tf.name_scope('accuracy'):
correct_predictions = tf.equal(self.predictions,
tf.argmax(self.input_y, 1))
self.accuracy = tf.reduce_mean(tf.cast(correct_predictions,
'float'),
name='accuracy')
def __init__(self, config, pretrained_embedding):
self._input = tf.placeholder(dtype=tf.int32,
shape=[None, config['num_steps']],
name='input')
self._target = tf.placeholder(dtype=tf.int32,
shape=[None],
name='target')
self.batch_size = config['batch_size']
self.num_steps = config['num_steps']
self.embed_size = config['embed_size']
self.size = config['hidden_size']
self._lr = config['lr']
self.num_classes = config['num_classes']
self.keep_prob = tf.Variable(config['keep_prob'], trainable=False)
self.combine_mode = config['combine_mode']
self.weight_decay = config['weight_decay']
with tf.device("/cpu:0"):
embedding = tf.Variable(pretrained_embedding, dtype=tf.float32)
inputs = tf.nn.embedding_lookup(embedding, self._input)
inputs = tf.nn.dropout(
inputs,
self.keep_prob,
noise_shape=[tf.shape(self._input)[0], 1, self.embed_size])
def lstm_cell(input_size):
return tf.contrib.rnn.LSTMCell(input_size,
forget_bias=0.0,
state_is_tuple=True,
reuse=tf.get_variable_scope().reuse)
def attn_cell(input_size):
return tf.contrib.rnn.DropoutWrapper(
lstm_cell(input_size),
output_keep_prob=config['keep_prob'],
variational_recurrent=True,
dtype=tf.float32)
cell = tf.contrib.rnn.MultiRNNCell(
[attn_cell(self.embed_size) for i in range(config['num_layers'])])
self._initial_state = cell.zero_state(
tf.shape(self._input)[0], tf.float32)
outputs = []
state = self._initial_state
with tf.variable_scope("RNN"):
for time_step in range(self.num_steps):
if time_step > 0: tf.get_variable_scope().reuse_variables()
(cell_output, state) = cell(inputs[:, time_step, :], state)
outputs.append(cell_output)
if self.combine_mode == 'mean':
outputs = tf.stack(outputs, axis=1)
outputs = tf.reduce_mean(outputs, axis=1)
outputs = tf.nn.dropout(outputs, self.keep_prob)
elif self.combine_mode == 'last':
outputs = outputs[-1]
outputs = tf.dropout(outputs, self.keep_prob)
outputs = tf.nn.dropout(outputs, self.keep_prob)
else:
outputs = tf.stack(outputs, axis=1)
outputs = tf.reduce_mean(outputs, axis=1)
outputs = tf.nn.dropout(outputs, self.keep_prob)
softmax_w = tf.get_variable("softmax_w", [self.size, self.num_classes],
dtype=tf.float32)
softmax_b = tf.get_variable("softmax_b", [self.num_classes],
dtype=tf.float32)
logits = tf.matmul(outputs, softmax_w) + softmax_b
# update the cost variables
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=self._target, logits=logits)
self.l2_loss = sum(
tf.nn.l2_loss(tf_var) for tf_var in tf.trainable_variables())
self._cost = cost = tf.reduce_mean(
loss) + self.weight_decay * self.l2_loss
self._lr = tf.Variable(self._lr, trainable=False)
tvars = tf.trainable_variables()
grads, _ = tf.clip_by_global_norm(tf.gradients(cost, tvars),
config['max_grad_norm'])
optimizer = tf.train.AdamOptimizer(self._lr)
self._train_op = optimizer.apply_gradients(zip(grads, tvars))
self._new_lr = tf.placeholder(tf.float32,
shape=[],
name="new_learning_rate")
self._lr_update = tf.assign(self._lr, self._new_lr)
self.predicted_class = tf.cast(
tf.argmax(tf.nn.softmax(logits), axis=-1), tf.int32)
def IDCNN_layer(self, model_inputs, name="IDCNN_layer"):
"""IDCNN layer
shape of input after expand = [batch, in_height, in_width, in_channels]
in_height：一句话为1，in_width：seq length，in_channels：embedding dim
shape of filter = [filter_height, filter_width, in_channels, out_channels]
in_channels：embedding dim，out_channels：number of filters
关于 tf.nn.atrous_conv2d(value,filters,rate,padding,name=None）
value： [batch, height, width, channels]这样的shape，
[batch_size, sentence高度(1), sentence宽度(length), sentence通道数(embedding dim)]
filters：[filter_height, filter_width, channels, out_channels]
[卷积核的高度，卷积核的宽度，图像通道数，卷积核个数]
rate：int型的正数，空洞卷积默认stride=1
rate参数，它定义为我们在输入图像上卷积时的采样间隔，相当于卷积核中穿插了（rate-1）数量的“0”，
这样做卷积时就相当于对原图像的采样间隔变大了。rate=1时，就没有0插入，相当于普通卷积。
padding： string，”SAME”,”VALID”其中之一，决定不同边缘填充方式。
“VALID”，返回[batch,height-(filter_height + (filter_height - 1) * (rate - 1))+1,
width-(filter_width + (filter_width - 1) * (rate - 1))+1,out_channels]的Tensor
“SAME”，返回[batch, height, width, out_channels]的Tensor
:param model_inputs: [batch_size, num_steps, emb_size]
:return: [batch_size, num_steps, totalChannels]
"""
model_inputs = tf.expand_dims(model_inputs, 1) # 化为图像相同维度进行处理
with tf.variable_scope(name):
# init filter weights
shape = [1, self.filter_width, self.embedding_dim, self.num_filter]
weights = tf.get_variable("idcnn_filter",
shape=shape,
initializer=self.initializer)
# first cnn layer
cnn_out = tf.nn.conv2d(model_inputs,
weights,
strides=[1,1,1,1],
padding="SAME",
use_cudnn_on_gpu=True,
name="first cnn layer")
# dilate cnn layers
reuseFlag = True if self.dropout == 0.0 else False
eachModuleOut = []
totalChannels = 0
for i in range(self.num_sublayers): # dilate cnn modules
for j in range(len(self.dilate_rate)): # dilate layers of one module
# reuse first layer, or when dropout rate is 0.0
with tf.variable_scope("atrous-conv-layer-%d" % i,
reuse=True if (reuseFlag or i > 0) else False):
w = tf.get_variable(
"weights",
shape=[1, self.filter_width, self.num_filter, self.num_filter],
initializer=tf.contrib.layers.xavier_initializer())
b = tf.get_variable("bias", shape=[self.num_filter])
atrous_conv_out = tf.nn.atrous_conv2d(cnn_out,
w,
rate=self.dilate_rate[j],
padding="SAME")
conv = tf.nn.bias_add(atrous_conv_out, b)
conv = tf.nn.relu(conv)
# record every sub modules` outputs
if j == (len(self.dilate_rate) - 1):
eachModuleOut.append(conv)
# iterate cnn inputs
cnn_out = conv
totalChannels = self.num_filter * self.num_sublayers
# aggregate outputs
cnn_outs = tf.concat(values=eachModuleOut, axis=3)
cnn_outs = tf.dropout(cnn_outs, 1 - self.dropout)
cnn_outs = tf.squeeze(cnn_outs, [1]) # expanded dim
cnn_outs = tf.reshape(cnn_outs, [-1, totalChannels])
self.cnn_out_channels = totalChannels
return cnn_outs
def build(self,
rgb_input,
train=False,
num_classes=20,
random_init_fc8=False,
debug=False):
"""
:param rgb_input is a image batch tensor, [ None x height x width x
num_channels ], The shape is how images are loaded.
"""
with tf.name_scope("Processing"):
ch_r, ch_g, ch_b = tf.split(rgb_input, 3, axis=3)
bgr = tf.concat(
[ch_b - VGG_MEAN_B, ch_g - VGG_MEAN_G, ch_r - VGG_MEAN_R],
axis=3)
if debug:
bgr = tf.Print(bgr, [tf.shape(bgr)],
message="Shape of input image: ",
summarize=4,
first_n=1)
# VGG convolutional
do_debug = True
self.conv1_1 = self._conv_layer(bgr, "conv1_1", False)
self.conv1_2 = self._conv_layer(self.conv1_1, "conv1_2", False)
self.pool1 = self._max_pool(self.conv1_2, "pool1", do_debug)
self.conv2_1 = self._conv_layer(self.pool1, "conv2_1", False)
self.conv2_2 = self._conv_layer(self.conv2_1, "conv2_2", False)
self.pool2 = self._max_pool(self.conv2_2, "pool2", do_debug)
self.conv3_1 = self._conv_layer(self.pool2, "conv3_1", False)
self.conv3_2 = self._conv_layer(self.conv3_1, "conv3_2", False)
self.conv3_3 = self._conv_layer(self.conv3_2, "conv3_3", False)
self.pool3 = self._max_pool(self.conv3_3, "pool3", do_debug)
self.conv4_1 = self._conv_layer(self.pool3, "conv4_1", False)
self.conv4_2 = self._conv_layer(self.conv4_1, "conv4_2", False)
self.conv4_3 = self._conv_layer(self.conv4_2, "conv4_3", False)
self.pool4 = self._max_pool(self.conv4_3, "pool4", do_debug)
self.conv5_1 = self._conv_layer(self.pool4, "conv5_1", False)
self.conv5_2 = self._conv_layer(self.conv5_1, "conv5_2", False)
self.conv5_3 = self._conv_layer(self.conv5_2, "conv5_3", False)
self.pool5 = self._max_pool(self.conv5_3, "pool5", do_debug)
self.fc6 = self._fc_layer(self.pool5,
"fc6",
do_relu=True,
debug=do_debug)
if train:
self.fc6 = tf.dropout(self.fc6, 0.5)
self.fc7 = self._fc_layer(self.fc6,
"fc7",
do_relu=True,
debug=do_debug)
if train:
self.fc7 = tf.dropout(self.fc7, 0.5)
self.fc8 = self._fc_layer(self.fc7,
"fc8",
do_relu=False,
num_classes=num_classes,
debug=do_debug)
pool4_shape = tf.shape(self.pool4)
self.upscore2 = self._upscore_layer(self.fc8,
"upscore2",
ksize=4,
stride=2,
num_classes=num_classes,
up_w=pool4_shape[2],
up_h=pool4_shape[1],
debug=do_debug)
self.score_pool4 = self._score_layer(self.pool4,
"score_pool4",
num_classes=num_classes,
random_weight_stddev=0.001)
self.fuse_pool4 = tf.add(self.upscore2, self.score_pool4)
input_shape = tf.shape(bgr)
self.upscore32 = self._upscore_layer(self.fuse_pool4,
"upscore32",
ksize=32,
stride=16,
num_classes=num_classes,
up_w=input_shape[2],
up_h=input_shape[1],
debug=do_debug)
return
